Name: Fundus3D
Author: WilloTwisper

Imported from WilloTwisper/Fundus3D (AGENTS.md). Install upstream with npx skills add WilloTwisper/Fundus3D. Copyright stays with the author.

Fundus3D — Project Context

Overview

Fundus3D converts a single 2D fundus (retinal) photograph into a textured 3D eye model suitable for visualization, simulation, and export.

The core idea: a schematic model of the human eye provides the 3D retinal surface geometry, and a camera model back-projects the 2D image pixels onto that surface through the eye's optics.

Schematic Eye Models

Three models are implemented, all based on published anatomical data:

Navarro (1985) — Default

Navarro R, Santamaría J, Bescós J. "Accommodation-dependent model of the human eye with aspherics." JOSA A, 2(8):1273-1281, 1985.

Aspheric corneal surfaces (conic constant ≈ -0.26)
Gradient-index lens
Clinically validated for visual optics research
Axial length ~24 mm (emmetropic), adjustable for myopia/hyperopia
Anterior segment length: ~12 mm

Simplified

Spherical approximation
Constant refractive index lens
Fast prototyping, less anatomically accurate

Custom

User-specified segment lengths
Useful for pathological eye shapes (high myopia, staphyloma)

Camera Models

Pinhole Camera

Standard perspective projection
Focal length derived from image size and FOV
Default FOV: 45° (typical fundus camera)

Widefield Camera

Equisolid-angle projection (f = R / 2sin(θ/2))
Suitable for widefield/ultra-widefield fundus imaging
Default FOV: 200°

Pipeline

1. Preprocessing
   - Detect fundus circle via Hough transform / thresholding
   - Create alpha mask (circular or from FOV mask image)
   - Generate texture atlas (fundus photo + generic retina fill + sclera)

2. Eye Model Generation
   - Build schematic eye geometry from parameters
   - Two mesh modes:
     a) Patch: retinal spherical cap matching camera FOV
     b) Full: complete eye with layers (sclera, cornea, iris, lens,
        retina, choroid, ciliary body, optic nerve, vitreous)

3. Camera Projection
   - Map image pixels → 3D rays through the eye
   - Compute UV coordinates for retinal vertices
   - Optional Snell refraction correction at corneal surface

4. Export
   - GLB (glTF 2.0 binary) with PBR materials
   - OBJ with MTL
   - Named anatomical layers for selective rendering

Datasets Used

ODIR-5K: 5000 paired left/right fundus images, preprocessed to 512×512
HRF: High-resolution fundus images (3504×2336) with FOV masks

Parameters

Parameter	Range	Meaning
`axial_length`	20–32 mm	Eye length (24=emmetropic, >26=myopic)
`equatorial_diameter`	22–28 mm	Width at equator
`fov_deg`	20–200°	Camera field of view
`mesh_resolution`	0–7	Subdivision level (higher = finer)

Key References

Navarro R et al. (1985) — Schematic eye model with aspherics
Atchison DA et al. (2004) — Eye shape changes with myopia (axial elongation with preserved equatorial diameter)
Equisolid-angle projection for widefield fundus cameras

Install Fundus3D and it becomes one more candidate the router can pick - when it fits.

A listing is a routing target with a manifest: what it does, which domains it covers, what it costs and who publishes it. Once installed it sits beside your own models and tools, is scored like any other candidate for each request, and shows up in the trace when it wins. Ratings come from workspaces that installed it, one per account.

{ "ocm": "1", "id": "willotwisper-fundus3d-fundus3d", "kind": "skill", "name": "Fundus3D", "description": "Fundus3D converts a single 2D fundus (retinal) photograph into a textured 3D eye model suitable for visualization, simulation, and export.", "publisher": "WilloTwisper", "version": "1.0.0", "capabilities": { "domains": [ "general" ], "tags": [ "skill-md", "github-agents" ], "languages": [ "en" ] }, "quality_prior": 0.6, "examples": [ "Fundus3D converts a single 2D fundus (retinal) photograph into a textured 3D eye model suitable for visualization, simulation, and export." ], "primary": false, "metadata": { "source": { "provider": "github-agents", "repository": "https://github.com/WilloTwisper/Fundus3D", "path": "AGENTS.md", "ref": "340c8e2b10853ba3466fe262eb24915c8b98cd56", "url": "https://github.com/WilloTwisper/Fundus3D/blob/340c8e2b10853ba3466fe262eb24915c8b98cd56/AGENTS.md", "key": "WilloTwisper/Fundus3D/AGENTS.md" } }, "instructions": "# Fundus3D — Project Context\n\n## Overview\n\nFundus3D converts a single 2D fundus (retinal) photograph into a textured 3D\neye model suitable for visualization, simulation, and export.\n\nThe core idea: a schematic model of the human eye provides the 3D retinal\nsurface geometry, and a camera model back-projects the 2D image pixels\nonto that surface through the eye's optics.\n\n## Schematic Eye Models\n\nThree models are implemented, all based on published anatomical data:\n\n### Navarro (1985) — Default\n*Navarro R, Santamaría J, Bescós J. \"Accommodation-dependent model of the\nhuman eye with aspherics.\" J", "cost": { "context_tokens": 739 } }

Fundus3D

About